Original strategy for page-walking was adjust stack pointer and then
touch pages in order. This kind of asks for double-fault, because
if touch fails, then signal will be delivered to frame above adjusted
stack pointer. But touching pages prior adjusting stack pointer would
upset valgrind. As compromise let's adjust stack pointer in pages,
touching top of the stack. This still asks for double-fault, but at
least prevents corruption of neighbour stack if allocation is to
overstep the guard page.
Also omit predict-non-taken hints as they reportedly trigger illegal
instructions in some VM setups.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Some OSes, *cough*-dows, insist on stack being "wired" to
physical memory in strictly sequential manner, i.e. if stack
allocation spans two pages, then reference to farmost one can
be punishable by SEGV. But page walking can do good even on
other OSes, because it guarantees that villain thread hits
the guard page before it can make damage to innocent one...
Reviewed-by: Rich Salz <rsalz@openssl.org>
and reorganize/harmonize post-conditions.
Additional hardening following on from CVE-2016-0702
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
At the same time remove miniscule bias in final subtraction.
Performance penalty varies from platform to platform, and even with
key length. For rsa2048 sign it was observed to be 4% for Sandy
Bridge and 7% on Broadwell.
CVE-2016-0702
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
bn_get_bits5 was overstepping array boundary by 1 byte. It was exclusively
read overstep and data could not have been used. The only potential problem
would be if array happens to end on the very edge of last accesible page.
Reviewed-by: Richard Levitte <levitte@openssl.org>
x86_64 platform. It targets specifically RSA1024 sign (using ideas
from http://eprint.iacr.org/2011/239) and adds more than 10% on most
platforms. Overall performance improvement relative to 1.0.0 is ~40%
in average, with best result of 54% on Westmere. Incidentally ~40%
is average improvement even for longer key lengths.
